Back to Guides
Enterprise Software 15 min read

Decoding AI fluency: what a trained team can do that an untrained one cannot

Decoding AI fluency: what a trained team can do that an untrained one cannot

AI fluency is the ability to work effectively, efficiently, ethically, and safely with AI systems. That definition comes from a specific place and it holds up well, but it will not tell you the thing you want to know. The question behind most “AI fluency meaning” searches is comparative: what will a trained team do that an untrained one cannot? That is a question about observable behavior, and it is the one every AI enablement budget is quietly underwriting. The answer is narrower, more concrete, and more modest than the frameworks make it sound.

What AI fluency means, and who defined it

The phrase has an author, which is unusual for business vocabulary and worth knowing before you buy anything described with it.

The Framework for AI Fluency was written by Rick Dakan, a professor at Ringling College of Art and Design, and Joseph Feller, a professor of information systems at University College Cork. They define AI fluency as “the ability to work effectively, efficiently, ethically, and safely within emerging modalities of Human-AI interaction.” Anthropic built a free public course on that framework, which is why most writing on the term traces back to these two academics whether or not the page says so.

The framework splits human-AI work into three modes: automation, where the model works from direct instructions; augmentation, where human and model “co-define and co-execute tasks in an iterative way”; and agency, where a person “configures AI to independently perform future tasks.” Across those modes sit four competencies, the 4Ds.

Competency What the framework says Where it applies
Delegation Selecting “the right AI tools and techniques” for the work Before you type anything
Description “Effectively describing a vision and/or tasks to prompt useful AI behaviors” While briefing
Discernment “Accurately assessing the usefulness of AI outputs” After the response arrives
Diligence “Taking responsibility and vouching for final products created using AI” Throughout, and at sign-off

This is a good model, and it is also the ceiling of what almost every page on the subject offers.

Why a definition of AI fluency decides nothing

Three problems with stopping at the framework.

Competency names are not behaviors. “Discernment” tells you that checking output matters. It does not tell you what a person who has it does on a Tuesday with a client summary in front of them, which is the only form the skill takes.

The framework was written for a different room. Its home is an art-and-design college and its examples are creative work. That is a legitimate origin and a poor fit for someone funding training for forty people in finance, operations, and sales. Nobody has done the translation.

Levels without a scale are decoration. Several vendor pages sell a ladder of fluency levels. Almost none says how you would place a specific employee on a rung using work your company already produces.

The fix is to stop asking what fluency is and start asking what it looks like, which is the argument behind treating access and capability as separate purchases.

Six capability differences you can see in the work

These are the differences that show up first when we run fluency sessions on a team’s own documents, ordered by how early they appear rather than by importance.

Untrained team Trained team
1. Choosing the task Sends whatever is in front of them, or avoids the tool for anything that matters Has a working sense of which tasks the model is good at and which it quietly fails
2. Briefing One or two sentences, no source material Attaches the real inputs, states the audience, names the constraints and the format
3. Checking Reads for tone, ships it Checks the three or four claims that would be expensive if wrong
4. Reuse Starts from nothing every time Keeps what worked and reuses it as a standing instruction
5. Stopping Pushes through six rounds of mediocre output Abandons after two and does the task by hand
6. Owning it “The AI wrote it” Signs their name to it and can defend every line

Choosing the task. The most expensive untrained habit is not bad prompting. It is applying the tool uniformly, to everything or to nothing. A trained team has calibrated intuitions: draft the first version of a recurring report, yes; produce a figure nobody will verify, no. This is the delegation competency, and it returns the most per hour of training because it removes work rather than adding it.

Briefing. Untrained use looks like search: type a request, expect the tool to supply the context. Trained use looks like handing work to a capable new hire. Here is last quarter’s version, here is who reads it, here is what cannot change, here is the format. Most of the quality difference people attribute to tool choice comes from this.

Checking. The trained behavior is not checking everything, which is slower than doing the work by hand. It is triage: a number, a name, a date, and a quoted policy are the four categories that go wrong, and fluent prose is the least reliable signal of accuracy there is.

Reuse. The clearest fork in the road. An untrained team’s hundredth use is no better than its first. A trained team accumulates: a saved brief for the monthly report, a standing instruction for client emails, a checked template for proposal sections. Within weeks, capability that compounds looks nothing like capability that resets.

Stopping. Knowing when to quit is a skill, and an untrained team does not have it, because the tool is agreeable and the next attempt always feels close. Trained people have a round limit and the confidence to use it.

Owning it. The diligence competency, which in practice means nobody sends something they could not defend in a meeting. It is also what makes a usage policy enforceable, as a rule people have practiced rather than read.

What the scored evidence says about trained and untrained teams

Most claims in this field come from surveys asking people how skilled they feel. One dataset tests instead.

Section’s AI Proficiency Report 2026 scored 5,026 US knowledge workers on hands-on tasks rather than self-reports, out of 100. Two findings matter here.

The first is the training delta. Workers at organizations that trained them on the agents and automations being deployed averaged 44 out of 100. At organizations that deployed without training, the average was 25. Clear, sanctioned tool access showed a similar spread, 43 against 25, and a formal strategy 43 against 26. Training is the single largest controllable variable in the dataset.

The second is the baseline, and it is bleaker than the marketing. Across the sample, 20.9% scored as novices and 73.5% as experimenters, leaving 5.5% at practitioner or expert level. On specific tasks, 21% could write an effective prompt, up from 7% in late 2025, and 3.8% wrote workable instructions for an AI assistant.

Two conclusions follow, and the second is the one vendors leave out.

The lever is real. Nothing else in the data moves capability by 1.8x, and it is a condition a company controls directly.

Forty-four out of a hundred is not fluency. It is competent intermediate work, and the trained cohort is better rather than good. Any proposal implying that a session or two produces a fluent organization is contradicted by the only scored sample available.

The usage numbers sharpen the gap. BCG’s AI at Work 2026 survey of close to 12,000 employees found 74% of frontline staff using AI daily or a few times a week, up 23 points in a year, while only 36% feel adequately upskilled. Three-quarters using it, one-third trained. That is why usage dashboards overstate what a team can do.

How to test AI fluency on your own documents this week

A test beats a framework, and this one costs an afternoon.

  1. Pick three real tasks. One recurring document, one piece of analysis, one piece of client-facing communication. Real work from this month, not exercises.
  2. Give eight to ten people ninety minutes. Same tasks, their own tools, no coaching. Collect the submissions and the prompts. The prompts are the more informative artifact.
  3. Score each one to five on four things. Was the right task given to the tool. Was the brief adequate. Were the checkable claims checked. Would the author defend it in front of the client. These map to the 4Ds, which is the point.
  4. Keep the artifacts. They are your baseline, and without one you cannot claim improvement later.
  5. Retest eight to twelve weeks after training, on different tasks of the same type. Earlier measures recall; later loses the attribution.

Two failure modes to avoid. A confidence survey is not a test and correlates poorly with scored performance. A certificate is not a test either: a completed course, including the free one from Anthropic, records attendance and nothing about whether behavior changed. That distinction between what a team knows and what it demonstrably does is the whole content of an AI-fluent organization.

What AI enablement costs, and what the capability delta is worth

Market ranges, not quotes. Nobody ranking for this vocabulary publishes a number, which is its own signal. Enablement services are also priced by component rather than as one product, so compare line items rather than totals.

Component Market range What you are paying for
Facilitated LLM fluency workshop $2,000–$15,000 Hands-on sessions on your own documents, scaled by headcount and session count
Capability baseline and retest Often bundled The before-and-after measurement. A baseline with no retest is half a purchase
Custom automation project $25,000–$150,000 A separate project after capability exists, not part of enablement

Set that against what you already spend. A blended $25 per seat per month across 150 seats is $45,000 a year in access with no capability line attached. A workshop program sits at a fraction of that, and it is the comparison your CFO will make anyway.

The return shows up in time that already exists. BCG found 42% of regular AI users saving eight hours a week, with higher figures in marketing, IT, and HR, while 66% receive limited or no guidance on what to do with that time and more than half do not reinvest it in strategic work. Hours are being freed and then absorbed. Fluency training decides whether those hours become capacity or evaporate, which is a more defensible case than a productivity percentage nobody can audit. The components that produce it are set out in what an enablement program contains.

How long AI enablement takes to move a team

Short, if you measure the right thing.

BCG’s 2025 edition found regular usage sharply higher among employees receiving at least five hours of training with access to in-person coaching. Five hours is a useful floor: below it you are running an announcement. It is a threshold, not a curriculum.

A realistic schedule to a defensible number is about ninety days. The baseline takes days, training is measured in sessions rather than months, and the retest sits eight to twelve weeks after the last session. Anything promising a transformed company inside that window is selling a different product, and the sequencing that makes ninety days work is the subject of the longer program guide.

One caution on scope. Fluency training moves what people can do with the tools you already own. It does not fix a missing data platform, write your governance regime, or build custom agents. Bundling those into a fluency engagement is the most common way a budget triples without capability moving.

If you want a read on where your team sits against the six behaviors above, we run a free AI-readiness call that works through your current licenses, what your staff produces today, and what a defensible first program looks like at your headcount.

Frequently asked questions

What does AI fluency mean?

AI fluency is the ability to work effectively, efficiently, ethically, and safely with AI systems, a definition from the Framework for AI Fluency by Rick Dakan and Joseph Feller. In practice it is four things: choosing which tasks to give a model, briefing it properly, judging what comes back, and owning the result.

What is the difference between AI literacy and AI fluency?

Literacy is understanding, fluency is applied skill. Someone AI-literate knows a model can be confidently wrong. Someone AI-fluent hands it the right task, briefs it with real source material, checks the claims that matter, and signs their name to the output. Literacy is testable with a quiz; fluency only with work.

What are the 4Ds of AI fluency?

Delegation, Description, Discernment, and Diligence: deciding what to hand to the model, communicating the task and its context, evaluating the output critically, and taking responsibility for the final product. They come from the Dakan and Feller framework and are taught in Anthropic’s free course built on it.

Is AI fluency the same as prompt engineering?

No, and the conflation is expensive. Prompting is one competency of four. Most of the value is created before the prompt, in choosing the task, and after it, in verification and ownership. A program that teaches prompt templates and stops has covered a quarter of the subject.

How do you measure AI fluency?

With task-based assessment on real work. Give people tasks from their own jobs, collect the output and the prompts, and score both against a short rubric covering task choice, brief quality, verification, and ownership. Section’s 2026 study tested hands-on tasks rather than self-reports because confidence and demonstrated skill diverge sharply.

What percentage of employees are AI fluent?

In Section’s scored sample of 5,026 US knowledge workers, 5.5% reached practitioner or expert level, 73.5% were experimenters and 20.9% novices. On individual tasks the figures are lower: 21% could write an effective prompt and 3.8% wrote workable instructions for an AI assistant.

Does AI fluency require technical skills or coding?

No. It requires judgment about tasks, clear written communication, and the discipline to verify. The people who improve fastest in sessions are often strong writers and careful reviewers rather than technical staff, because three of the four competencies are editorial skills.

How long does it take to become AI fluent?

Measurable improvement takes weeks, not months. BCG found regular usage sharply higher above five hours of training with coaching access, and ninety days is enough to produce a before-and-after number. Reaching the top tier is a longer project and is not what most teams need.

Can we just use a free course instead of paying for training?

Partly. Anthropic’s free course teaches the framework well and is worth assigning. What it cannot do is work on your documents, score your people, or tell you whether behavior changed. Free material plus an internal baseline and retest is a legitimate low-budget program. Free material alone is reading.

Is AI fluency legally required anywhere?

Something adjacent is. Article 4 of the EU AI Act, in force since 2 February 2025, requires providers and deployers to take measures supporting AI literacy among staff operating AI systems on their behalf. It sets no curriculum, no exam, and no certification requirement, so it creates a documentation duty for companies in scope rather than a fluency bar.

Key takeaways

  • AI fluency has a specific definition and a specific author: the ability to work effectively, efficiently, ethically, and safely with AI systems, from Dakan and Feller’s framework, organized around delegation, description, discernment, and diligence.
  • The definition is not the useful part. The behavioral delta is, and it shows up in six places: which tasks go to the model, how they are briefed, what gets checked, whether anything is reused, when people stop, and who owns the output.
  • Scored testing of 5,026 US knowledge workers found 44 out of 100 where employers trained people on the systems they deployed, against 25 where they did not. No other variable in that dataset moves capability as much.
  • The same data caps the claim. Trained teams are better, not fluent, and 5.5% of the sample reached practitioner level. Treat a vendor promising a fluent organization in a quarter accordingly.
  • You can baseline this yourself in an afternoon: three real tasks, ninety minutes, keep the prompts, score task choice, briefing, verification, and ownership, then retest eight to twelve weeks later.

Last Updated: Nov 18, 2026

AW

Arthur Wandzel

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles